AI agent or automation? Pay for the simplest tool that works

One quick way to waste money on technology is to use an AI agent for a workflow problem. The other is to use fixed rules for work full of judgement calls. The first mistake is easy to make right now, because "AI agent" is the pitch you hear most.

This article helps you choose; it isn't a pitch. Many requests that reach us as "AI agent" projects suit Power Automate or a rule in your ERP better. When that's true, we say so, because we don't sell AI where a simple workflow will do.

Automation follows your rules; an agent handles the judgement calls

Rules-based automation runs fixed steps when something triggers it, so the same input always gives the same output. For example, an invoice arrives and the flow reads its supplier and amount. It matches the invoice to the purchase order, approves it if within tolerance, then posts it to the ERP.

An AI agent reads the situation, decides what to do, acts, then uses the result to decide its next step. It can handle inputs that don't fit a set pattern, and make something like a judgement call. Its output can vary with context nobody planned for.

So the question is simple. Does your problem need judgement on unclear cases, or reliable running of a known sequence?

If you can write down every rule, and the rules rarely change, you probably need an automation, not an agent. For well-defined work, it is faster, cheaper and more reliable.

Choose automation when you can write every rule down

The process is well defined. Every decision can be written as a rule. "If the invoice is within 5% of the purchase order, approve it. Otherwise, send it to the finance manager."

The inputs are structured. Data arrives in a predictable format, such as system exports, web forms or database records. Not free-text emails, handwritten forms or PDFs with changing layouts.

Exceptions are few and easy to handle. The automation handles most cases. The few exceptions go to a person, with the details attached.

The rules are stable. The logic rarely changes. When it does, someone non-technical can often update it without a rebuild.

Volume is the bottleneck. The work is manual because there is too much of it, not because each case needs hard thinking.

Power Automate, and the workflow rules in many ERPs and CRMs, suit this kind of work. They are cheap to run, quick to build and easier to maintain than custom code.

If your problem fits this description, start here. It is what our Process automation service builds.

Choose an agent when the work needs judgement on messy inputs

The inputs are messy or varied. Think free-text questions, documents in many formats, or customer emails phrased a hundred ways. Or the same kind of input needs different handling, depending on context.

The work needs interpretation. The question isn't "does this match a rule?" It's "what is this asking for, and what's the right response for this customer and product?"

Many cases need a person. Our team's rule of thumb: if more than about 40% of cases need a person's judgement, consider an agent. Below that, rules plus an exception queue usually cope.

The work takes several steps of reasoning. The agent gathers information from several sources, weighs it, then recommends or acts. The steps aren't fixed; they depend on what it finds.

The value is in the judgement, not the volume. Tender responses are a good example. Every tender asks for different things, and the best answer draws on the right past work.

This is the work our Custom AI agents service is for. A person approves any action that matters.

An agent costs more to build and run, so it has to earn that cost

An automation for a well-defined process, such as invoice matching and approval routing, can often be built in weeks. With a tool like Power Automate, the build is modest and licences are cheap. Upkeep is low.

An agent for messy, judgement-heavy work, such as tender drafts or purchasing decisions, takes longer and costs more. It needs monitoring, tuning and upkeep of the documents it draws on. You can't set it and forget it.

In return, it does what an automation can't: it handles unclear cases at volume. On the right tasks, its quality can approach a skilled person's.

So an agent makes economic sense when one of these is true:

• The volume is high enough that the time saved outweighs the yearly build and running cost.
• Quality matters so much that today's output is costing you work, such as tender responses that sound generic.

In our team's experience, that tends to mean hundreds of cases a month, not dozens. Or it means a quality gap that wins or loses jobs.

Five questions that tell you which one you need

Ask these in order for any process:

1. Can you write every decision as a fixed rule? → Yes: automate. No: go to question 2.
2. Does the work involve free text or documents that vary? → Yes: it likely needs an agent.
3. What share of cases needs a person's judgement today? → Our team's rule of thumb: under 20%, automate and send exceptions to a person. Over 40%, consider an agent.
4. How many cases are there a month? → Our team's rule of thumb: under 50, an agent probably won't pay for itself yet. Over 200, it usually will.
5. Does each answer need information from several sources? → Yes: agent. No: automation may be enough.

The answers usually point one way. If they don't, start with automation: it's cheaper and faster. If it falls short, you'll see where, and an agent becomes a clear second project.

A worked example: an "AI agent" brief that needed a simple workflow

This is an illustration, not a real project. Say a company asks for "an AI agent to process supplier invoices". Look closer, and the process is fully defined.

The invoices come from a few suppliers in a fixed format. Approval works on a tolerance, and routing has two steps. Nothing is unclear, and nothing is free text.

So the right build is a Power Automate flow, not an agent. It costs less, goes live sooner, and sends the odd exception to a person with the details attached. An agent here would be impressive and unnecessary.